Machine Learning in Perioperative Medicine: A Comparative Review of Predictive, Causal, and Foundation Model Approaches in Surgical Data Science

Abstract

This review aimed to clarify the methodological landscape of machine learning (ML) in perioperative medicine by comparing three major paradigms: supervised predictive models, causal inference frameworks, and foundation models. It addressed the central research question of which ML approach is most appropriate for specific perioperative clinical scenarios and under what data conditions. A narrative review was conducted using peer-reviewed literature from PubMed, Scopus, and Web of Science (2015–2026). Studies were included if they reported quantitative ML performance in surgical or anesthesia contexts and addressed at least one of the three paradigms. Evidence was synthesized to compare methodological characteristics, performance, interpretability, and clinical applicability. Supervised models (e.g., XGBoost, Random Forest) dominate current practice, demonstrating strong predictive performance but lacking causal interpretability. Causal ML approaches, including meta-learners and doubly robust estimators, effectively capture treatment heterogeneity but rely on strong assumptions and complex validation. The TabPFN foundation model performs well on small datasets without tuning but shows limitations in calibration and regression tasks. Across all paradigms, external validation is limited, and interpretability and generalizability remain key challenges. No single ML paradigm is universally optimal for perioperative applications. Method selection should align with the clinical question and data structure. Future progress depends on multicenter validation, integration of paradigms, and development of regulatory-compliant, interpretable clinical decision-support systems.